Loading...
1/* SPDX-License-Identifier: GPL-2.0-only */
2/*
3 * Copyright (C) 2013 Red Hat
4 * Author: Rob Clark <robdclark@gmail.com>
5 */
6
7#ifndef __MSM_RINGBUFFER_H__
8#define __MSM_RINGBUFFER_H__
9
10#include "msm_drv.h"
11
12#define rbmemptr(ring, member) \
13 ((ring)->memptrs_iova + offsetof(struct msm_rbmemptrs, member))
14
15#define rbmemptr_stats(ring, index, member) \
16 (rbmemptr((ring), stats) + \
17 ((index) * sizeof(struct msm_gpu_submit_stats)) + \
18 offsetof(struct msm_gpu_submit_stats, member))
19
20struct msm_gpu_submit_stats {
21 u64 cpcycles_start;
22 u64 cpcycles_end;
23 u64 alwayson_start;
24 u64 alwayson_end;
25};
26
27#define MSM_GPU_SUBMIT_STATS_COUNT 64
28
29struct msm_rbmemptrs {
30 volatile uint32_t rptr;
31 volatile uint32_t fence;
32
33 volatile struct msm_gpu_submit_stats stats[MSM_GPU_SUBMIT_STATS_COUNT];
34};
35
36struct msm_ringbuffer {
37 struct msm_gpu *gpu;
38 int id;
39 struct drm_gem_object *bo;
40 uint32_t *start, *end, *cur, *next;
41 struct list_head submits;
42 uint64_t iova;
43 uint32_t seqno;
44 uint32_t hangcheck_fence;
45 struct msm_rbmemptrs *memptrs;
46 uint64_t memptrs_iova;
47 struct msm_fence_context *fctx;
48 spinlock_t lock;
49};
50
51struct msm_ringbuffer *msm_ringbuffer_new(struct msm_gpu *gpu, int id,
52 void *memptrs, uint64_t memptrs_iova);
53void msm_ringbuffer_destroy(struct msm_ringbuffer *ring);
54
55/* ringbuffer helpers (the parts that are same for a3xx/a2xx/z180..) */
56
57static inline void
58OUT_RING(struct msm_ringbuffer *ring, uint32_t data)
59{
60 /*
61 * ring->next points to the current command being written - it won't be
62 * committed as ring->cur until the flush
63 */
64 if (ring->next == ring->end)
65 ring->next = ring->start;
66 *(ring->next++) = data;
67}
68
69#endif /* __MSM_RINGBUFFER_H__ */
1/* SPDX-License-Identifier: GPL-2.0-only */
2/*
3 * Copyright (C) 2013 Red Hat
4 * Author: Rob Clark <robdclark@gmail.com>
5 */
6
7#ifndef __MSM_RINGBUFFER_H__
8#define __MSM_RINGBUFFER_H__
9
10#include "drm/gpu_scheduler.h"
11#include "msm_drv.h"
12
13#define rbmemptr(ring, member) \
14 ((ring)->memptrs_iova + offsetof(struct msm_rbmemptrs, member))
15
16#define rbmemptr_stats(ring, index, member) \
17 (rbmemptr((ring), stats) + \
18 ((index) * sizeof(struct msm_gpu_submit_stats)) + \
19 offsetof(struct msm_gpu_submit_stats, member))
20
21struct msm_gpu_submit_stats {
22 u64 cpcycles_start;
23 u64 cpcycles_end;
24 u64 alwayson_start;
25 u64 alwayson_end;
26};
27
28#define MSM_GPU_SUBMIT_STATS_COUNT 64
29
30struct msm_rbmemptrs {
31 volatile uint32_t rptr;
32 volatile uint32_t fence;
33
34 volatile struct msm_gpu_submit_stats stats[MSM_GPU_SUBMIT_STATS_COUNT];
35 volatile u64 ttbr0;
36};
37
38struct msm_cp_state {
39 uint64_t ib1_base, ib2_base;
40 uint32_t ib1_rem, ib2_rem;
41};
42
43struct msm_ringbuffer {
44 struct msm_gpu *gpu;
45 int id;
46 struct drm_gem_object *bo;
47 uint32_t *start, *end, *cur, *next;
48
49 /*
50 * The job scheduler for this ring.
51 */
52 struct drm_gpu_scheduler sched;
53
54 /*
55 * List of in-flight submits on this ring. Protected by submit_lock.
56 *
57 * Currently just submits that are already written into the ring, not
58 * submits that are still in drm_gpu_scheduler's queues. At a later
59 * step we could probably move to letting drm_gpu_scheduler manage
60 * hangcheck detection and keep track of submit jobs that are in-
61 * flight.
62 */
63 struct list_head submits;
64 spinlock_t submit_lock;
65
66 uint64_t iova;
67 uint32_t hangcheck_fence;
68 struct msm_rbmemptrs *memptrs;
69 uint64_t memptrs_iova;
70 struct msm_fence_context *fctx;
71
72 /**
73 * hangcheck_progress_retries:
74 *
75 * The number of extra hangcheck duration cycles that we have given
76 * due to it appearing that the GPU is making forward progress.
77 *
78 * For GPU generations which support progress detection (see.
79 * msm_gpu_funcs::progress()), if the GPU appears to be making progress
80 * (ie. the CP has advanced in the command stream, we'll allow up to
81 * DRM_MSM_HANGCHECK_PROGRESS_RETRIES expirations of the hangcheck timer
82 * before killing the job. But to detect progress we need two sample
83 * points, so the duration of the hangcheck timer is halved. In other
84 * words we'll let the submit run for up to:
85 *
86 * (DRM_MSM_HANGCHECK_DEFAULT_PERIOD / 2) * (DRM_MSM_HANGCHECK_PROGRESS_RETRIES + 1)
87 */
88 int hangcheck_progress_retries;
89
90 /**
91 * last_cp_state: The state of the CP at the last call to gpu->progress()
92 */
93 struct msm_cp_state last_cp_state;
94
95 /*
96 * preempt_lock protects preemption and serializes wptr updates against
97 * preemption. Can be aquired from irq context.
98 */
99 spinlock_t preempt_lock;
100};
101
102struct msm_ringbuffer *msm_ringbuffer_new(struct msm_gpu *gpu, int id,
103 void *memptrs, uint64_t memptrs_iova);
104void msm_ringbuffer_destroy(struct msm_ringbuffer *ring);
105
106/* ringbuffer helpers (the parts that are same for a3xx/a2xx/z180..) */
107
108static inline void
109OUT_RING(struct msm_ringbuffer *ring, uint32_t data)
110{
111 /*
112 * ring->next points to the current command being written - it won't be
113 * committed as ring->cur until the flush
114 */
115 if (ring->next == ring->end)
116 ring->next = ring->start;
117 *(ring->next++) = data;
118}
119
120#endif /* __MSM_RINGBUFFER_H__ */